The world's most secretive intelligence alliance doesn't often pick up a megaphone. Yesterday it did — and the message was that the cyber-threat landscape is about to change, fast, because of AI.

"Not years — months"

The Five Eyes — the signals-intelligence agencies of the US, UK, Canada, Australia and New Zealand — issued a rare joint statement on 22 June warning that frontier AI models are poised to "fundamentally transform both offensive and defensive cyber capabilities." Their timeframe was blunt: "The timeline is not years, it is months."

It's worth being precise about what they did and didn't say, because the headlines have run ahead of the agencies. The Guardian's framing — that AI "could take down governments and business" — is the newspaper's compression, not the agencies' words. What the officials (including the NSA's cyber chief and the acting head of the US's CISA) actually warned is narrower but still serious: that frontier models will sharply lower the barrier to sophisticated attacks — automating the discovery of software flaws, generating malware, accelerating exploit development, and powering large-scale, convincing social engineering. They named the model class driving the concern: Anthropic's Fable 5 and OpenAI's Daybreak, expected to be broadly available within the year.

Their advice was practical and, notably, defensive: shrink your attack surface, patch faster, treat ageing legacy systems as a strategic liability, tighten identity and access controls, and assume you'll be breached. The line that captures it: "Adversaries are already using AI to move faster and more effectively. Defenders must do the same." For a community that usually communicates in classified channels, going public is itself the signal.

An AI law firm wins in court — with an asterisk

A neat illustration of AI's quieter, real-world creep: an English court case has been won with the help of Garfield AI, described as the UK's first regulator-approved AI law firm. A freelance HR consultant, chasing £7,000 in unpaid fees, recovered the full sum at Wandsworth County Court — with Garfield handling the pre-action letters, the filings, the witness statements and the trial bundle for a fee of around £400.

The honest framing matters here too: the AI didn't argue the case. A human barrister still ran the courtroom advocacy, which he stressed remained an essentially human exercise. What's genuinely new is that all the preparation — the slow, expensive paperwork that prices many people out of justice — was done by a regulated AI service. It's thought to be a first, and it points at where AI may bite first in professional services: not replacing the expert in the room, but doing everything that leads up to it.

Compute's growing power bill

The other thread that keeps thickening is energy. Chevron and Microsoft announced a 20-year deal on 22 June for Chevron to build a natural-gas power plant in West Texas — "Project Kilby," near Pecos — dedicated to feeding a Microsoft AI data centre. It will ramp to 2.67 gigawatts, enough to power more than half a million Texas homes, with electricity flowing by 2028.

The number to sit with is that one: a single AI site now warrants its own multi-gigawatt power station. It also sharpens a tension we keep returning to — TechCrunch notes the plant is projected to emit over 13 million tons of CO₂ a year, awkward against Microsoft's pledge to be carbon-negative by 2030. The race for compute and the race to decarbonise are increasingly pulling in opposite directions, and AI is where they collide.

Also worth a glance

A 3-billion-parameter model called VibeThinker, fine-tuned in China and released open under an MIT licence, is making the rounds for matching far larger models on narrow maths and coding benchmarks. It's a genuinely interesting efficiency result — but a specialist one, by its own creators' admission, which falls behind on broader knowledge tasks. File it under "small models are getting cleverer at specific things," not "frontier in your pocket."

A quiet Tuesday it isn't. The throughline: AI's capabilities, costs and risks are all compounding at once — and this week the people whose job is to worry about the risks decided to say so out loud.